Latest / The Tech Matchmaker / AI Token Costs, Agent Sprawl & Enterprise AI Strategy | Krista
Transcript
- Gina: Good morning everybody. Today we're catching back up with Aaron Symbolik the AI Orchestration Consultant at Krista.ai. Aaron, welcome back to the Tech Matchmaker. Aaron Symbolik: Thank you so much, Gina. It's always great to see your smiling face. Gina: It's been about a year since we last chatted and clearly a lot has happened over at Krista. So for anyone who missed our last episode, can you give us a quick reminder who you are and what Krista.ai does? Aaron Symbolik: Yeah, I'm an a I'm what I call probably an AI orchestration or AI consultant. and that doesn't mean I'm on or independent, but I work directly for a company called Krista AI out of Dallas, Texas, by a really brilliant founder, John Michelsen. And so we're based out of Dallas. I'm in South Orange County, California, is where I'm located. And I eat, sleep and breathe, everything machine learning, artificial intelligence. large language models, but Krista AI itself, we are an orchestration platform. So we use a lot of A 17 plus types of AI capabilities to build workflows on top of companies' existing technology. So instead of people doing a lot of the manual work and people becoming, you know, bottlenecks in a process, we leverage machines to move at machine speed and make outcomes a lot easier to achieve for businesses. That's really it in nutshell. Gina: I love it. And I'm gonna ask you to expand a little bit on your elevator pitch as far as what problems you guys solve for. And then I wanna dive into what's new. Aaron Symbolik: Yeah, sure. So I I think there's some general issues that we're seeing out there, but problems that we see particularly is optimizing sales, you know, the a sales force, right? Not Salesforce, the product, but sales force is and how you're getting rid of a lot of the administrative and mundane work they deal with, or it could be customer support internally. There's downstream and upstream processes from customer support. When someone calls in, they have, we need a copy of this, or we need to do this sort of function. Krista actually takes action from a person asking, you know, to do something or a system trigger. triggers Krista to run a workflow. So an example of that might be somebody's requesting a purchase order and it has rules around it. Well, as you know, if you use a large language model, like some of the celebrity or frontier models, those use what we call probabilistic logic. Two plus two equals probably gonna be four. We use some of that, but we also use deterministic logic. And so Krista builds these orchestrations with guardrails, security guardrails to make sure it doesn't that 10 out of 10 times if you're asking a purchase order, it does it accurately 10 out of 10 times. And it doesn't just make up the 10th time. If it doesn't know something or needs something information to fit finish that 10th PO, it's going to go to a person. So we always believe we call it person loop, but I call it wisdom in loop. 'Cause it is truly Gina: Mm. Aaron Symbolik: the wisdom of a person to to iterate on a process. And you all never want to get rid of the human. people are scared that AI is replacing humans, maybe displacing in some positions or jobs, but Krista's job is not to replace the human, it's actually to make the human more powerful, iterate on a process and take all that mundane work that humans don't need to do and move at machine speed. That's essentially what we're doing. I know it sounds very holistic. Gina: Yeah, I I love that. I think one of the other big problems, and I know you and I were chatting about this the other day when we were talking about a customer, but is the token expense management side. And and that's Aaron Symbolik: Yes. Gina: something that I I think as we discuss, like almost every company that uses an AI model is struggling with right now. So talk to us about how Krista's solving for that. Aaron Symbolik: So I'm gonna give you a good lead into that because right now the big conversation, the three problems that we're hearing out there are exactly that's the first one. Our token costs are out of control, or we just saw double what we expected this to be, and our costs are outweighing the value, what we anticipated the value we would get from that particular tool. The other one are the security concerns that I kind of was explaining, right? Is it do we have proper guardrails around these agents? We just saw a recent thing where open AI. duplicated itself and went into another software application like that Gina: Yes. Aaron Symbolik: we can protect against you can use these tools and protect against things like that. The other one, the big issues still, and I was saying this a year ago, and it's even compounding more, is agent sprawl. I don't want to say we can't, we're not an agentic platform, but we're the antithesis of building a bunch of agents, individual productivity tools running around your organization. So those are the three big issues. So the the million dollar question is what is a mature AI AI strategy? And where do you get started? Because it could be very overwhelming if I sat and did a presentation for a potential customer showing them all these different things that we can do, but there's got to be a starting point. And here at Krista, we talk about really three what I call gateways that just get started and get value in week one. One of those is let's start building the Krista brain. The Krista brain is at any organization you have dynamic information. Those are the conversations Gina, like we're having right now. It's dynamic, it's happening in real time. How do we capture that information? Gina, you're very experienced. You have a lot of knowledge in that brain. So if we could capture that somehow and hone that in for you know newer, younger folks that are coming to the workplace, it would be great if we can hone in on some of that wisdom and knowledge that you've gained over the years. and so you wanna build that brain on dynamic knowledge, you want to build it on system knowledge, and you wanna build it in all those areas like SharePoint where people are housing that. If you want to build a good AI strategy, start with building the brain, but then guess what? We're using agents right now and we're or costs are getting out of control. So if you were starting with Krista, we'd say, we can do some really interesting use cases for you, but let's start by controlling your existing agents. So if you're using Claude agents, co-work, for example, and your costs are getting out of control, we have something called LLM access governance. So we govern against your costs, spend limits, give you visibility into all the tools your employees are using, but how much are they spending on those and actually set up shut off budgets? Where once it hits a certain budget, we gotta shut that down because we're now losing value. We've hit the curve and now we're on the downside where we're not getting as much value out of it. So you have that's called critical mass. Now that we understand what the problems that people are facing, instead of trying to moonshot and build a bunch of use cases for clients, we see a great starting point is building the enterprise brain. human knowledge, dynam dynamic knowledge, tr we called tribal knowledge, right? For people that have been to their companies, system knowledge, that's the tech stack they have, and of course, wherever they keep all their SOPs, SharePoint, all that data. You start to build that enterprise brain. But immediately what we're finding, that problem of people using large language models or agents using cowork and their costs are getting out of control. Once they reach critical mass, then they start losing money because they're spending too much, then you're losing the value of those agents. So we have a way, a starting point with Krista to use what's called LLM access governance. That gives you complete visibility into all the tools your employees are using, how much they're spending, and setting spend limits. And even setting the roles as far as who can access what within those tools, all the while setting up the security around them so they're not getting your IP or company's IPs leaving that and going into the LLMs. So you get complete visibility and control over all your costs so that you're getting to that critical mass and not losing money. So that would be a great starting point. And the way our customers see it, we're getting value immediately on the AI tools that we're already using. And all the while we can actually run multiple models within this governance to say, hey, by the way, you're using this, you're using OpenAI for this. We actually found that Claude is less expensive and a better model to use. So we can actually match all the different models and tell them what's a better model from a cost and functionality standpoint. In many cases, Gina, we're seeing they want to use Krista because we can show them we can do it at a fraction of the cost and way more with more guardrails around how that deterministic logic using. Instead of using all probabilistic logic, meaning two plus two equals, well, it got it said five. Well, it's not five. We know that. Gina: Yeah. Aaron Symbolik: Krista's not going to get it wrong. It's going to loop in humans where those celebrity frontier models are not going to loop in humans. Ours will, because it we have IP around that, that if it gets it wrong or it thinks it's going to get it wrong, it's going to ask somebody with tribal knowledge to get it right. So you can iterate on the process and make it smarter as you go on. Gina: Well, and I think, Aaron, you've hit on an issue everybody listening can relate to, which is using Claude or using OpenAI, and it just assumes, right? And then you're wasting tokens on things that you're like, where did that slop come from? Right. So yeah, that's a pretty incredible cost savings for a customer. I mean, to be able to have that Aaron Symbolik: Yeah. Yeah. Gina: recommendation of use this, not this. talk a little bit about the security concerns. Aaron Symbolik: Yeah, so if you think about the you know, traditionally we're using, well, now we're using APIs, right? To do a lot of integrations. And you know, there's a lot of employees that are using things on their own for company purposes. There's concerns there. We just saw the again OpenAI leaving and creating an agent and hacking into somebody else's software. We saw that. We want to put guardrails around, you know, that we want to run you, you can run those agents through Krista. Put the security necessary around it because we were a SOC2 type two ISO 2701, but we look at those guards and we set rules, like a framework for how that agent should operate and never operate outside of that. So it's like basically putting train training wheels on your agents to make sure they don't get out of control and they stay in their lane, right? Or putting them on Gina: Yeah. Aaron Symbolik: the train tracks and they're gonna go down this path, but they're never gonna veer off into a different path. That's probably the best analogy how you want to do it. So where people start is. You know, it's great. We can get the LL access control to kind of see where we're spending and what we're using, but now we can use what's called MCP server with Krista to then put guardrails around it and put the security layer that you guys would want. You're not going to go into your business. It's like hiring employees and not giving them any training and telling what the rules are and letting them do whatever they want rogue. That's essentially what people are doing right now with their agents. We would Gina: Yeah. Aaron Symbolik: we we give them the training, we give them the guardrails, we tell them the framework. how they're supposed to operate in a business. And by the way, if you want to build something, Krista, you can do that too. But let's just start with getting control of what your employees are using, get control of your cost, put security layers around that. That's the mature AI strategy. And I haven't found anybody that would debate me on that and say that's not a mature way to start using AI. Gina: I love that. Now let's go back to kind of the basis of of how AI works, which is the documentation. And you mentioned SharePoint a little bit earlier, but you guys have a document understanding agent. You did an update to that a little bit earlier this year, I believe. Can you Aaron Symbolik: Yeah. Gina: walk us through what that actually does and why it mattered enough to build it? Aaron Symbolik: Well, it matters because traditional OCR is we need to cr let's just take something as simple as creating a purchase order and or we and we need to get artifacts off a document. Traditional OCR says it's optical zoom, right? So we're gonna go up two inches, over three inches. There should be a PO number there. Grab that. There should be this piece of information, a a dollar figure there. Grab that. And that's what traditional OCR is. What Artificial intelligence allows us to do, even the celebrity models do it, how they don't always get it right, is you can take unstructured documents and data and pull that information off there and then create a purchase order, for example. The difference in Krista, though, is that we train Krista on the types of documents you're going to get, what they should look like. And by the way, if you don't get all this information correct and you haven't got the artifacts we told you to. You're now going to go to a h human in the loop or wisdom in the loop and ask that person, I'm missing one more piece of information that I need in order to place this purchase order. As opposed to if I run it through one of the celebrity models, it's just gonna probably it's designed to give an answer. It may just go ahead and create that purchase order. And you could have out of a hundred, you might get 10 wrong, but that might be a quarter million dollars that you just lost in in inventory that you just sent out that you shouldn't have. That's why you need guardrails Gina: Yeah. Aaron Symbolik: and security around these types of consequential. you know, costly decisions that that your computers are making for you. Gina: Love that. Aaron, we talked about it last time. The token expense management is definitely number one in my my books. But what what other features or functionalities set Krista apart from the competitors in the market? Because I know when we first met, you know, orchestration wasn't commonly used, but now agentic AI and the word orchestration is used, it's showing up everywhere. So what are the differentiators for Krista? And and as I said, that token expense management is. Clearly a big one. Aaron Symbolik: The I hear it in every conversation, by the way. and I I I I hate to admit this, I wouldn't say it's an accidental product, but it's something we were already doing for ourselves internally and we were doing it for our clients internally, meaning it was only facing us so that we can make sure that they weren't and then we're like, why aren't we letting our clients see this? 'Cause people they started asking that question, like, we need we're we're using other tools and we need this ability to see that. And that's why we said, my gosh, we have a product sitting in our back. in our back room here that we haven't even thought about productizing and now we're productizing it and it's actually the first crank of the wheel for our clients. But the biggest differentiator, because there's all I hear this all the time. we've we talked to another company that's just like you. And I look them up and I found out they're a consulting firm. And I'm like, that's great, they're consulting firm. So you're telling me they're gonna go out and bolt on different products and try to get what we are. I haven't found anything quite like us that does everything that we do. So velocity is the name of the game. How quick can we build something with security and guardrails? Because I can go into co it co-work and maybe build a personal productivity to go out and find clients for me, right? Let's just say it does that and it it gives me information. I can build that pretty quickly. That's what clients want. We don't hold the keys to the kingdom and we're not Oz, the great Oz behind a a curtain. We actually, the new feature that we're offering, and this isn't new. You've always been able to access the background Krista and see your workflows. But you don't have to be a code writer. To write your orchestration, your processes, we don't hold that, hold you hostage to our product. We actually can flip it, give you the back end, and now we've just designed it to where you can talk in natural language. Krista becomes the data scientist and will build your workflows. So you're not taking a mouse pointing and clicking at drop-downs, drag and drop and building your workflow. Krista actually will have an iterative conversation with you, back and forth, questions to know what she has to build. And what systems that you have that she needs to integrate with, whether it's your email, your Salesforce, Zendesk, your ERP system, any of those tools, she can integrate. Jira. You know, we have a client where they want to do password resets. She can integrate with Jira and do password resets for you. But you can actually build those yourself using natural language rather than using drop downs, coding, data scientists. So if I were to say our biggest win and biggest infiltrator, it's our velocity and how fast we can actually deploy. use cases for our clients. Gina: Love that. Who's the ideal customer for you guys right now? Like what does a company look like right before they realize, my goodness, we need Krista? Aaron Symbolik: I think any any customer they could find value quickly with with any kind of customer support, high volume inquiries, whether it be an agent assist internally or whether it be outward facing towards the client. I think any business oriented white collar large company where they have a lot of processes they're doing over and over, making hundreds of decisions a day. typically I would say clients that are a little bit on the probably larger side. And the only reason I say that is not that we're not a fit for a let's say a smaller company that has a lot of redundancy issues. It's just that naturally you find that larger companies have a lot more human friction, red tape, and bottlenecks in their processes. So we s we tend to be a very good fit for a client. When I say larger, I say maybe, you know, 500 employees plus manufacturing, could be legal, could be I said white collar, anything that's the sales focused or CX focused customer experience. Anything around that's really a good fit. But I haven't really come into many industries that come to me and don't have some sort of outcome they're trying to solve where it's there's a lot of redundancy and bottlenecks. but Gina: Yeah. Aaron Symbolik: those are kind of that's where we play it seem to be the to have the most conversations. Gina: Do you want to share a specific use case or customer? You don't have to give the name. over the past year we're that's getting some really incredible results from Krista. Aaron Symbolik: Yeah, I'd say we have a client. I mean, I could go on and talk about many clients, but we have a client where, you know, they they're a security company and they want the ability to basically get off the phone as quick as possible and get resolution to their clients. So we built something in the iframe of their CRM where they can actually answer questions. The Krista will identify the client, identify the equipment they have. Instead of going through all those FAQs and SOPs on what kind of equipment that is and how to solve their problem, Krista can quickly identify that, answer questions. So really what they want to avoid there is not only just speed to getting getting at resolution, but not having to send out new customer service calls. We also have another one where it's a utility company, people are having issues or high bills or questions about their invoicing, and Krista can integrate with their smart meter and see that there actually is a technical issue. And automatically go out and schedule service calls and then put things like stoppage on the bill so they're not getting invoiced until their problem is fixed. Can do a lot of smart things like that and then only involve human when there's exceptions to the rules. so maybe there's an exception of some sort that needs to take place so there would be human in the loop. So that's another example where we're t taking that front end CX, but doing it a smart way contextually with their existing tech stack. Because if you did that with a regular chatbot or a voice agent. it's not going to be have all those those deep integrations in their smart system and know it's smart enough to shut down the billing and then get schedule a call and do all that in one stream conversation. So we get in a lot of complex complex workflows, work streams, if you will. Gina: I love that. Is there anything, Aaron, that you'd like the audience to know that I haven't asked you yet? Aaron Symbolik: No, I don't think so. I think the biggest thing is don't try to moonshot and over engineer products. Start really simple, maybe with a smart meeting agent LM access control. Start with building your your brain so you can get a mature AI strategy. And then maybe pick one, you know, starting point that's not consequential and get started on a use case. That's really your best way to start with Krista. Or really any AI for that matter. Gina: Yeah. And now comes your favorite part, the rapid fire. Aaron Symbolik: okay. Gina: So what sound or noise do you love? Aaron Symbolik: Sound or noise do I love? I love hearing actually every day, I love hearing the waterfalls come out of my pool. Kinda is a Zen-like experience. Gina: I love that. And what noise or sound do you hate? Aaron Symbolik: I hate when people are chewing their food. Like I cannot stand it. Literally drives me up a wall. I have to leave the room. If anybody's chewing and their mouth opens once or they're chewing gum with their mouth open, that drives me absolute bonkers. Gina: And what profession other than your own would you like to attempt? Aaron Symbolik: Would I like to be if I wasn't doing that? Gina: Yeah, yeah. Aaron Symbolik: I'd like to be a trial attorney. Gina: I think you'd be phenomenal at that. And Aaron, if anybody wants to learn more about Krista, where should they go? Aaron Symbolik: Two places. They can go to our website, which is www.krista.ai or they could reach out to Gina and Gina will get us in touch with you. Yeah. Gina: Excellent. Thanks so much, Aaron. It's always great catching up with you. And Krista is clearly not standing still right now. And as we know, neither is this industry. So thanks so much for coming back and spending time with us. Can't wait to have you again because I know you guys are going to be changing the the industry shortly. So thank you. Aaron Symbolik: Awesome. Thank you so much, Gina. It's always a pleasure.